Domain intelligible models

Abstract

Mining biological information from rich "-omics" datasets is facilitated by organizing features into groups that are related to a biological phenomenon or clinical outcome. For example, microorganisms can be grouped based on a phylogenetic tree that depicts their similarities regarding genetic or physical characteristics. Here, we describe algorithms that incorporate auxiliary information in terms of groups of predictors and the relationships between them into the metagenome learning task to build intelligible models. In particular, our cost function guides the feature selection process using auxiliary information by requiring related groups of predictors to provide similar contributions to the final response. We apply the developed algorithms to a recently published dataset analyzing the effects of fecal microbiota transplantation (FMT) in order to identify factors that are associated with improved peripheral insulin sensitivity, leading to accurate predictions of the response to the FMT.